Skip to contentSkip to conversation

Fix MSR_IA32_PLATFORM_ID access for TDX guests

PATCH In review Binbin Wu 6 participants 16 messages Apr 28 – Apr 29

Subsystems (5)
lore
Share & tools
lore

Binbin Wu

author
#0 lore raw

cover Fix MSR_IA32_PLATFORM_ID access for TDX guests

Since commit d8630b67ca1e ("x86/cpu: Add platform ID to CPU info structure"), the Linux kernel reads MSR_IA32_PLATFORM_ID during CPU init. This causes unchecked MSR access errors when running as a TDX guest, because KVM doesn't include MSR_IA32_PLATFORM_ID in tdx_has_emulated_msr().

On KVM side, this patch set adds MSR_IA32_PLATFORM_ID to tdx_has_emulated_msr() so that KVM returns 0 for TDs reading this MSR, consistent with how KVM already handles it for normal VMs.

Optionally, in order to provide back-compatibility for newer Linux guests running on older KVM hosts without the fix, the second patch skips the MSR_IA32_PLATFORM_ID read entirely when Linux is running in a virtualized environment. This could be dropped if we think the fix on KVM side is sufficient.

Binbin Wu (2):
KVM: TDX: Allow TDs to read MSR_IA32_PLATFORM_ID
x86/cpu: Skip reading MSR_IA32_PLATFORM_ID in virtualized environment

 arch/x86/kernel/cpu/microcode/core.c     | 2 +-
 arch/x86/kernel/cpu/microcode/intel.c    | 4 ++++
 arch/x86/kernel/cpu/microcode/internal.h | 5 +++++
 arch/x86/kvm/vmx/tdx.c                   | 1 +
 4 files changed, 11 insertions(+), 1 deletion(-)

base-commit: 39704f00f747aba3144289870b5fd8ac230a9aaf

Signature · 1 line
2.46.0
author
#1 lore raw

PATCH 1/2 KVM: TDX: Allow TDs to read MSR_IA32_PLATFORM_ID

Add MSR_IA32_PLATFORM_ID to tdx_has_emulated_msr() so that TDs can read it.

Linux kernel reads MSR_IA32_PLATFORM_ID during init since commit d8630b67ca1e ("x86/cpu: Add platform ID to CPU info structure"). KVM already supports this MSR on read for normal VMs by returning 0. Without support for this MSR, TDs get unchecked MSR access errors.

    unchecked MSR access error: RDMSR from 0x17 at rIP: 0xffffffffba38d6fc (intel_get_platform_id+0x7c/0xb0)
    Call Trace:
     <TASK>
     ? early_init_intel+0x28/0x2c0
     ? early_cpu_init+0x9b/0x930
     ? setup_arch+0xbf/0xbb0
     ? _printk+0x6b/0x90
     ? start_kernel+0x7f/0xaa0
     ? x86_64_start_reservations+0x24/0x30
     ? x86_64_start_kernel+0xda/0xe0
     ? common_startup_64+0x13e/0x141
     </TASK>

Add MSR_IA32_PLATFORM_ID in tdx_has_emulated_msr() to allow TDs to read the MSR. MSR_IA32_PLATFORM_ID is read-only by hardware definition, i.e. KVM should never add it as writable, no need to add it in tdx_is_read_only_msr().

Fixes dd50294f3e3c ("KVM: TDX: Implement callbacks for MSR operations")
---
 arch/x86/kvm/vmx/tdx.c | 1 +
 1 file changed, 1 insertion(+)
arch/x86/kvm/vmx/tdx.c +1 -0
index 04ce321ebdf3..812ad99b11e5 100644
@@ -2094,6 +2094,7 @@ void tdx_get_exit_info(struct kvm_vcpu *vcpu, u32 *reason,
 bool tdx_has_emulated_msr(u32 index)
 {
 	switch (index) {
+	case MSR_IA32_PLATFORM_ID:
 	case MSR_IA32_UCODE_REV:
 	case MSR_IA32_ARCH_CAPABILITIES:
 	case MSR_IA32_POWER_CTL:
Signature · 1 line
2.46.0
author
#10 lore raw

PATCH 2/2 x86/cpu: Skip reading MSR_IA32_PLATFORM_ID in virtualized environment

In reply to Binbin Wu

The Linux kernel now reads MSR_IA32_PLATFORM_ID during CPU init. When running as a guest, if the underlying hypervisor does not emulate the MSR, the RDMSR from 0x17 in intel_get_platform_id() can trigger an unchecked MSR access during early boot.

    unchecked MSR access error: RDMSR from 0x17 at rIP: 0xffffffffba38d6fc (intel_get_platform_id+0x7c/0xb0)
    Call Trace:
     <TASK>
     ? early_init_intel+0x28/0x2c0
     ? early_cpu_init+0x9b/0x930
     ? setup_arch+0xbf/0xbb0
     ? _printk+0x6b/0x90
     ? start_kernel+0x7f/0xaa0
     ? x86_64_start_reservations+0x24/0x30
     ? x86_64_start_kernel+0xda/0xe0
     ? common_startup_64+0x13e/0x141
     </TASK>

Currently, KVM does not support guest reads of MSR_IA32_PLATFORM_ID for TDX. This should be fixed in the hypervisor, but for compatibility with newer guests running on older KVM hosts, skip reading MSR_IA32_PLATFORM_ID and return 0. It's meaningless to read MSR_IA32_PLATFORM_ID since guests are not allowed to do microcode update. cpu_has_old_microcode(), which uses platform ID for CPU match, also skips checking whether the microcode is old or not in a virtualized environment.

Since intel_get_platform_id() can be called early before cpuinfo_x86 is fully initialized, check whether it's running in a virtualized environment from CPUID and opportunistically add a helper.

Fixes d8630b67ca1ed ("x86/cpu: Add platform ID to CPU info structure")
---
 arch/x86/kernel/cpu/microcode/core.c     | 2 +-
 arch/x86/kernel/cpu/microcode/intel.c    | 4 ++++
 arch/x86/kernel/cpu/microcode/internal.h | 5 +++++
 3 files changed, 10 insertions(+), 1 deletion(-)
arch/x86/kernel/cpu/microcode/core.c +1 -1
index 651202e6fefb..ee204c8a90bf 100644
@@ -135,7 +135,7 @@ bool __init microcode_loader_disabled(void)
 	 * 3) Certain AMD patch levels are not allowed to be
 	 *    overwritten.
 	 */
-	hypervisor_present = native_cpuid_ecx(1) & BIT(31);
+	hypervisor_present = x86_cpuid_has_hypervisor();
 
 	if ((hypervisor_present && !IS_ENABLED(CONFIG_MICROCODE_DBG)) ||
 	    amd_check_current_patch_level())
arch/x86/kernel/cpu/microcode/intel.c +4 -0
index 37ac4afe0972..cb93e4ea410e 100644
@@ -147,6 +147,10 @@ u32 intel_get_platform_id(void)
 	if (intel_cpuid_vfm() <= INTEL_PENTIUM_II_KLAMATH)
 		return 0;
 
+	/* Don't try to read microcode bits when virtualized. */
+	if (x86_cpuid_has_hypervisor())
+		return 0;
+
 	/* get processor flags from MSR 0x17 */
 	native_rdmsr(MSR_IA32_PLATFORM_ID, val[0], val[1]);
 
arch/x86/kernel/cpu/microcode/internal.h +5 -0
index 3b93c0676b4f..0233e074d76b 100644
@@ -100,6 +100,11 @@ static inline unsigned int x86_cpuid_family(void)
 	return x86_family(eax);
 }
 
+static inline bool x86_cpuid_has_hypervisor(void)
+{
+	return native_cpuid_ecx(1) & BIT(31);
+}
+
 extern bool force_minrev;
 
 #ifdef CONFIG_CPU_SUP_AMD
Signature · 1 line
2.46.0
Reviewed-by on v1 1/2
#2 lore raw

Re: [PATCH 1/2] KVM: TDX: Allow TDs to read MSR_IA32_PLATFORM_ID

In reply to Binbin Wu
Quoted 21 lines · Binbin Wu

Add MSR_IA32_PLATFORM_ID to tdx_has_emulated_msr() so that TDs can read it.

Linux kernel reads MSR_IA32_PLATFORM_ID during init since commit d8630b67ca1e ("x86/cpu: Add platform ID to CPU info structure"). KVM already supports this MSR on read for normal VMs by returning 0. Without support for this MSR, TDs get unchecked MSR access errors.

     unchecked MSR access error: RDMSR from 0x17 at rIP: 0xffffffffba38d6fc (intel_get_platform_id+0x7c/0xb0)
     Call Trace:
      <TASK>
      ? early_init_intel+0x28/0x2c0
      ? early_cpu_init+0x9b/0x930
      ? setup_arch+0xbf/0xbb0
      ? _printk+0x6b/0x90
      ? start_kernel+0x7f/0xaa0
      ? x86_64_start_reservations+0x24/0x30
      ? x86_64_start_kernel+0xda/0xe0
      ? common_startup_64+0x13e/0x141
      </TASK>
Quoted 8 lines

Add MSR_IA32_PLATFORM_ID in tdx_has_emulated_msr() to allow TDs to read the MSR. MSR_IA32_PLATFORM_ID is read-only by hardware definition, i.e. KVM should never add it as writable, no need to add it in tdx_is_read_only_msr().

Fixes dd50294f3e3c ("KVM: TDX: Implement callbacks for MSR operations")

Nit: It would be helpful to add the info that read on this MSR from TDs always triggers #VE so that Linux guest will request emulation from KVM.

Quoted 16 lines
---
  arch/x86/kvm/vmx/tdx.c | 1 +
  1 file changed, 1 insertion(+)
arch/x86/kvm/vmx/tdx.c +1 -0
index 04ce321ebdf3..812ad99b11e5 100644
@@ -2094,6 +2094,7 @@ void tdx_get_exit_info(struct kvm_vcpu *vcpu, u32 *reason,
  bool tdx_has_emulated_msr(u32 index)
  {
  	switch (index) {
+	case MSR_IA32_PLATFORM_ID:
  	case MSR_IA32_UCODE_REV:
  	case MSR_IA32_ARCH_CAPABILITIES:
  	case MSR_IA32_POWER_CTL:
#11 lore raw

Re: [PATCH 2/2] x86/cpu: Skip reading MSR_IA32_PLATFORM_ID in virtualized environment

In reply to Binbin Wu
Quoted 57 lines · Binbin Wu

The Linux kernel now reads MSR_IA32_PLATFORM_ID during CPU init. When running as a guest, if the underlying hypervisor does not emulate the MSR, the RDMSR from 0x17 in intel_get_platform_id() can trigger an unchecked MSR access during early boot.

     unchecked MSR access error: RDMSR from 0x17 at rIP: 0xffffffffba38d6fc (intel_get_platform_id+0x7c/0xb0)
     Call Trace:
      <TASK>
      ? early_init_intel+0x28/0x2c0
      ? early_cpu_init+0x9b/0x930
      ? setup_arch+0xbf/0xbb0
      ? _printk+0x6b/0x90
      ? start_kernel+0x7f/0xaa0
      ? x86_64_start_reservations+0x24/0x30
      ? x86_64_start_kernel+0xda/0xe0
      ? common_startup_64+0x13e/0x141
      </TASK>

Currently, KVM does not support guest reads of MSR_IA32_PLATFORM_ID for TDX. This should be fixed in the hypervisor, but for compatibility with newer guests running on older KVM hosts, skip reading MSR_IA32_PLATFORM_ID and return 0. It's meaningless to read MSR_IA32_PLATFORM_ID since guests are not allowed to do microcode update. cpu_has_old_microcode(), which uses platform ID for CPU match, also skips checking whether the microcode is old or not in a virtualized environment.

Since intel_get_platform_id() can be called early before cpuinfo_x86 is fully initialized, check whether it's running in a virtualized environment from CPUID and opportunistically add a helper.

Fixes d8630b67ca1ed ("x86/cpu: Add platform ID to CPU info structure")
---
  arch/x86/kernel/cpu/microcode/core.c     | 2 +-
  arch/x86/kernel/cpu/microcode/intel.c    | 4 ++++
  arch/x86/kernel/cpu/microcode/internal.h | 5 +++++
  3 files changed, 10 insertions(+), 1 deletion(-)
arch/x86/kernel/cpu/microcode/core.c +1 -1
index 651202e6fefb..ee204c8a90bf 100644
@@ -135,7 +135,7 @@ bool __init microcode_loader_disabled(void)
  	 * 3) Certain AMD patch levels are not allowed to be
  	 *    overwritten.
  	 */
-	hypervisor_present = native_cpuid_ecx(1) & BIT(31);
+	hypervisor_present = x86_cpuid_has_hypervisor();
 
  	if ((hypervisor_present && !IS_ENABLED(CONFIG_MICROCODE_DBG)) ||
  	    amd_check_current_patch_level())
arch/x86/kernel/cpu/microcode/intel.c +0 -0
index 37ac4afe0972..cb93e4ea410e 100644

arch/x86/kernel/cpu/microcode/intel.c u32 intel_get_platform_id(void) · line 150

Quoted 4 lines
arch/x86/kernel/cpu/microcode/intel.c +1 -0
@@ -147,6 +147,10 @@ u32 intel_get_platform_id(void)
  	if (intel_cpuid_vfm() <= INTEL_PENTIUM_II_KLAMATH)
  		return 0;
 
+	/* Don't try to read microcode bits when virtualized. */

The platform ID is not only used by microcode update. I don't think calling them microcode bits is proper. It's also stashed in struct cpuinfo_x86::intel_platform_id and used in x86_match_cpu() as a factor as the generic FMS (Family, Model, Stepping) bits.

We skip things due to hypervisor is set usually when the thing is known unable to be virtualized. From the perspective of microcode update, it's OK to skip reading it due to hypervisor bit. However, from the perspective of generic platform ID, it seems not that reasonable to skip reading it due to hypervisor bit. Especially KVM has supported this MSR for normal VMs.

So how about using the safe version of RDMSR, if we want the enhancement?

Quoted 22 lines
arch/x86/kernel/cpu/microcode/intel.c +3 -0
+	if (x86_cpuid_has_hypervisor())
+		return 0;
+
  	/* get processor flags from MSR 0x17 */
  	native_rdmsr(MSR_IA32_PLATFORM_ID, val[0], val[1]);
 
arch/x86/kernel/cpu/microcode/internal.h +5 -0
index 3b93c0676b4f..0233e074d76b 100644
@@ -100,6 +100,11 @@ static inline unsigned int x86_cpuid_family(void)
  	return x86_family(eax);
  }
 
+static inline bool x86_cpuid_has_hypervisor(void)
+{
+	return native_cpuid_ecx(1) & BIT(31);
+}
+
  extern bool force_minrev;
 
  #ifdef CONFIG_CPU_SUP_AMD
author
#12 lore raw

Re: [PATCH 2/2] x86/cpu: Skip reading MSR_IA32_PLATFORM_ID in virtualized environment

On 4/28/2026 2:01 PM, Xiaoyao Li wrote:
[...]

Quoted 10 lines
Quoted 8 lines
arch/x86/kernel/cpu/microcode/intel.c +0 -0
index 37ac4afe0972..cb93e4ea410e 100644
@@ -147,6 +147,10 @@ u32 intel_get_platform_id(void)
       if (intel_cpuid_vfm() <= INTEL_PENTIUM_II_KLAMATH)
           return 0;

  +    /* Don't try to read microcode bits when virtualized. */

The platform ID is not only used by microcode update. I don't think calling them microcode bits is proper. It's also stashed in struct cpuinfo_x86::intel_platform_id and used in x86_match_cpu() as a factor as the generic FMS (Family, Model, Stepping) bits.

It's still related to microcode though.
The platform id is used to detect whether the microcode is old or not in this case.

Quoted 2 lines

We skip things due to hypervisor is set usually when the thing is known unable to be virtualized. From the perspective of microcode update, it's OK to skip reading it due to hypervisor bit. However, from the perspective of generic platform ID, it seems not that reasonable to skip reading it due to hypervisor bit. Especially KVM has supported this MSR for normal VMs.

Considering the microcode related code, Linux kernel does special handling when running as a guest, microcode_loader_disabled() sets dis_ucode_ldr and cpu_has_old_microcode() also skips to check whether the microcode is old if it's running in a virtualized environment. I.e. the special handling for virtualization make it meaningless to read the MSR.

Quoted 2 lines

So how about using the safe version of RDMSR, if we want the enhancement?

I am open to this.
Let's wait to see if the change on guest side is wanted or not.

Quoted 22 lines
Quoted 20 lines

+    if (x86_cpuid_has_hypervisor())
+        return 0;
+
      /* get processor flags from MSR 0x17 */
      native_rdmsr(MSR_IA32_PLATFORM_ID, val[0], val[1]);
  diff --git a/arch/x86/kernel/cpu/microcode/internal.h b/arch/x86/kernel/cpu/microcode/internal.h index 3b93c0676b4f..0233e074d76b 100644

arch/x86/kernel/cpu/microcode/internal.h +4 -0
@@ -100,6 +100,11 @@ static inline unsigned int x86_cpuid_family(void)
       return x86_family(eax);
   }
   +static inline bool x86_cpuid_has_hypervisor(void)
+{
+    return native_cpuid_ecx(1) & BIT(31);
+}
+
   extern bool force_minrev;
     #ifdef CONFIG_CPU_SUP_AMD
#3 lore raw

Re: [PATCH 1/2] KVM: TDX: Allow TDs to read MSR_IA32_PLATFORM_ID

In reply to Binbin Wu
Quoted 38 lines · Binbin Wu

Add MSR_IA32_PLATFORM_ID to tdx_has_emulated_msr() so that TDs can read it.

Linux kernel reads MSR_IA32_PLATFORM_ID during init since commit d8630b67ca1e ("x86/cpu: Add platform ID to CPU info structure"). KVM already supports this MSR on read for normal VMs by returning 0. Without support for this MSR, TDs get unchecked MSR access errors.

   unchecked MSR access error: RDMSR from 0x17 at rIP: 0xffffffffba38d6fc (intel_get_platform_id+0x7c/0xb0)
   Call Trace:
    <TASK>
    ? early_init_intel+0x28/0x2c0
    ? early_cpu_init+0x9b/0x930
    ? setup_arch+0xbf/0xbb0
    ? _printk+0x6b/0x90
    ? start_kernel+0x7f/0xaa0
    ? x86_64_start_reservations+0x24/0x30
    ? x86_64_start_kernel+0xda/0xe0
    ? common_startup_64+0x13e/0x141
    </TASK>

Add MSR_IA32_PLATFORM_ID in tdx_has_emulated_msr() to allow TDs to read the MSR. MSR_IA32_PLATFORM_ID is read-only by hardware definition, i.e. KVM should never add it as writable, no need to add it in tdx_is_read_only_msr().

Fixes dd50294f3e3c ("KVM: TDX: Implement callbacks for MSR operations")

---
arch/x86/kvm/vmx/tdx.c | 1 +

1 file changed, 1 insertion(+)
arch/x86/kvm/vmx/tdx.c +0 -0
index 04ce321ebdf3..812ad99b11e5 100644

arch/x86/kvm/vmx/tdx.c void tdx_get_exit_info(struct kvm_vcpu *vcpu, u32 *reason, · line 2100

Quoted 7 lines
arch/x86/kvm/vmx/tdx.c +1 -0
@@ -2094,6 +2094,7 @@ void tdx_get_exit_info(struct kvm_vcpu *vcpu, u32 *reason,
 bool tdx_has_emulated_msr(u32 index)
 {
 	switch (index) {
+	case MSR_IA32_PLATFORM_ID:
 	case MSR_IA32_UCODE_REV:
 	case MSR_IA32_ARCH_CAPABILITIES:
 	case MSR_IA32_POWER_CTL:

This patch looks good to me. But the rule for which MSRs should be emulated by KVM for TDX is not very clear to me.

Maybe we can document the rule here, if there is one. That would make it much easier to tell whether future issues like this are guest regressions or missing KVM handling.

#4 lore raw

Re: [PATCH 1/2] KVM: TDX: Allow TDs to read MSR_IA32_PLATFORM_ID

Quoted 34 lines · Chao Gao
Quoted 33 lines · Binbin Wu

Add MSR_IA32_PLATFORM_ID to tdx_has_emulated_msr() so that TDs can read it.

Linux kernel reads MSR_IA32_PLATFORM_ID during init since commit d8630b67ca1e ("x86/cpu: Add platform ID to CPU info structure"). KVM already supports this MSR on read for normal VMs by returning 0. Without support for this MSR, TDs get unchecked MSR access errors.

   unchecked MSR access error: RDMSR from 0x17 at rIP: 0xffffffffba38d6fc (intel_get_platform_id+0x7c/0xb0)
   Call Trace:
    <TASK>
    ? early_init_intel+0x28/0x2c0
    ? early_cpu_init+0x9b/0x930
    ? setup_arch+0xbf/0xbb0
    ? _printk+0x6b/0x90
    ? start_kernel+0x7f/0xaa0
    ? x86_64_start_reservations+0x24/0x30
    ? x86_64_start_kernel+0xda/0xe0
    ? common_startup_64+0x13e/0x141
    </TASK>

Add MSR_IA32_PLATFORM_ID in tdx_has_emulated_msr() to allow TDs to read the MSR. MSR_IA32_PLATFORM_ID is read-only by hardware definition, i.e. KVM should never add it as writable, no need to add it in tdx_is_read_only_msr().

Fixes dd50294f3e3c ("KVM: TDX: Implement callbacks for MSR operations")

---
arch/x86/kvm/vmx/tdx.c | 1 +

1 file changed, 1 insertion(+)
Quoted 15 lines

arch/x86/kvm/vmx/tdx.c void tdx_get_exit_info(struct kvm_vcpu *vcpu, u32 *reason, · line 2100

Quoted 12 lines
arch/x86/kvm/vmx/tdx.c +1 -0
index 04ce321ebdf3..812ad99b11e5 100644
@@ -2094,6 +2094,7 @@ void tdx_get_exit_info(struct kvm_vcpu *vcpu, u32 *reason,
 bool tdx_has_emulated_msr(u32 index)
 {
 	switch (index) {
+	case MSR_IA32_PLATFORM_ID:
 	case MSR_IA32_UCODE_REV:
 	case MSR_IA32_ARCH_CAPABILITIES:
 	case MSR_IA32_POWER_CTL:

This patch looks good to me. But the rule for which MSRs should be emulated by KVM for TDX is not very clear to me.

I would strongly prefer to not take this patch, and instead fix the guest. This isn't some latent/pre-existing guest behavior, it's brand new functionality. I.e. Linux-as-a-TDX-guest broke itself.

More importantly from a guest security perspective, consuming MSR_IA32_PLATFORM_ID is actively dangerous. E.g. it could allow the untrusted host to steer the guest's behavior with respect to feature detection and enablement.

And once KVM allows reads from MSR_IA32_PLATFORM_ID, there's no going back. E.g. if the TDX-Module wants to emulate MSR_IA32_PLATFORM_ID, because there's an actual *need* to do so, then we're going to have a (minor) mess with KVM's ABI.

Quoted 3 lines

Maybe we can document the rule here, if there is one. That would make it much easier to tell whether future issues like this are guest regressions or missing KVM handling.

#5 lore raw

Re: [PATCH 1/2] KVM: TDX: Allow TDs to read MSR_IA32_PLATFORM_ID

Quoted 11 lines · Sean Christopherson
Quoted 2 lines

This patch looks good to me. But the rule for which MSRs should be emulated by KVM for TDX is not very clear to me.

I would strongly prefer to not take this patch, and instead fix the guest. This isn't some latent/pre-existing guest behavior, it's brand new functionality. I.e. Linux-as-a-TDX-guest broke itself.

More importantly from a guest security perspective, consuming MSR_IA32_PLATFORM_ID is actively dangerous.  E.g. it could allow the untrusted host to steer the guest's behavior with respect to feature detection and enablement.

I feel like you could make similar cases about a lot the VMM controlled MSRs.

Quoted 5 lines

And once KVM allows reads from MSR_IA32_PLATFORM_ID, there's no going back. E.g. if the TDX-Module wants to emulate MSR_IA32_PLATFORM_ID, because there's an actual *need* to do so, then we're going to have a (minor) mess with KVM's ABI.

This is a problem with the bare metal KVM behavior too, and it's just super old?

For TDX, hmm. I guess the standard thing to do in order to avoid creating a KVM ABI problems is just match the arch behavior. But for TDs, it is a very special type of VM. The special TDX guest things can't work on bare metal. Furthermore, guest opt ins can change what arch is even supposed to be virtualized. So the normal KVM default thing to do doesn't always fit.

So instead we will just virtualize as little as possible to keep Linux guest running? Ok.

Quoted 4 lines
Quoted 4 lines

Maybe we can document the rule here, if there is one. That would make it much easier to tell whether future issues like this are guest regressions or missing KVM handling.

Speaking in general now about TDX guest ABI, not about this specific MSR. I agree it would be good to have some story about how this is supposed to work and evolve. Similar to the CPUID bit stuff we are trying to clean up.

I'm also remembering the debate about what to do about the MTRR bit being forced on, but not being able to apply the KVM solution (which got ripped out anyway) to the S-EPT. Or the guest maxpfn thing, where we got the TDX module to add CPUID bit support we needed. By default, we can try to make it so the TDX module will, or will let us, faithfully match the normal x86 arch bit behavior.

#6 lore raw

Re: [PATCH 1/2] KVM: TDX: Allow TDs to read MSR_IA32_PLATFORM_ID

Quoted 8 lines · Rick P Edgecombe
Quoted 7 lines · Sean Christopherson
Quoted 2 lines

This patch looks good to me. But the rule for which MSRs should be emulated by KVM for TDX is not very clear to me.

I would strongly prefer to not take this patch, and instead fix the guest. This isn't some latent/pre-existing guest behavior, it's brand new functionality. I.e. Linux-as-a-TDX-guest broke itself.

Quoted 6 lines
Quoted 4 lines

More importantly from a guest security perspective, consuming MSR_IA32_PLATFORM_ID is actively dangerous.  E.g. it could allow the untrusted host to steer the guest's behavior with respect to feature detection and enablement.

I feel like you could make similar cases about a lot the VMM controlled MSRs.

Oh, for sure. And far worse (yet another reminder that I need to refresh this series).

https://lore.kernel.org/all/[email protected]

Quoted 6 lines
Quoted 4 lines

And once KVM allows reads from MSR_IA32_PLATFORM_ID, there's no going back. E.g. if the TDX-Module wants to emulate MSR_IA32_PLATFORM_ID, because there's an actual *need* to do so, then we're going to have a (minor) mess with KVM's ABI.

This is a problem with the bare metal KVM behavior too, and it's just super old?

No? Oh, I see what you're asking. I'm mostly concerned about the host side of things. The problem with TDX is that a TDX-Module update could effectively change KVM's behavior, i.e. if the TDX-Module decides it needs to emulate PLATFORM_ID for whatever reason. So not only would KVM need to enumerate to userspace that the MSR is supported/emulated for TDX guests, KVM would also need to differentiate between emulated by KVM and emulated by the TDX-Module.

Quoted 8 lines

For TDX, hmm. I guess the standard thing to do in order to avoid creating a KVM ABI problems is just match the arch behavior. But for TDs, it is a very special type of VM. The special TDX guest things can't work on bare metal. Furthermore, guest opt ins can change what arch is even supposed to be virtualized. So the normal KVM default thing to do doesn't always fit.

So instead we will just virtualize as little as possible to keep Linux guest running? Ok.

Yeah, and the sequence of events matters. Most of KVM's half-hearted emulation of random MSRs exists because KVM needed to be able to run existing kernels. But once KVM (and other hypervisors) existed, kernels learned to run as guests (and hardware vendors largely did a better job of explicitly enumerating MSRs and whatnot), and so the need to throw hacks into KVM mostly went away.

Quoted 13 lines
Quoted 3 lines
Quoted 3 lines

Maybe we can document the rule here, if there is one. That would make it much easier to tell whether future issues like this are guest regressions or missing KVM handling.

Speaking in general now about TDX guest ABI, not about this specific MSR. I agree it would be good to have some story about how this is supposed to work and evolve. Similar to the CPUID bit stuff we are trying to clean up.

I'm also remembering the debate about what to do about the MTRR bit being forced on, but not being able to apply the KVM solution (which got ripped out anyway) to the S-EPT. Or the guest maxpfn thing, where we got the TDX module to add CPUID bit support we needed. By default, we can try to make it so the TDX module will, or will let us, faithfully match the normal x86 arch bit behavior.

#8 lore raw

Re: [PATCH 1/2] KVM: TDX: Allow TDs to read MSR_IA32_PLATFORM_ID

In reply to Binbin Wu
Quoted 4 lines · Binbin Wu

Linux kernel reads MSR_IA32_PLATFORM_ID during init since commit d8630b67ca1e ("x86/cpu: Add platform ID to CPU info structure"). KVM already supports this MSR on read for normal VMs by returning 0. Without support for this MSR, TDs get unchecked MSR access errors.

NAK from me on this.

The platform ID is used for one thing and one thing only: microcode updates. Those updates are solely the domain of the bare-metal OS.

The (guest) kernel code that's even trying to touch this MSR is buggy and insane. We need to turn that code it off when X86_FEATURE_HYPERVISOR==1. There's already a patch floating around to do that.

Please don't add more smoke and mirrors.

#13 lore raw

Re: [PATCH 2/2] x86/cpu: Skip reading MSR_IA32_PLATFORM_ID in virtualized environment

In reply to Xiaoyao Li
Quoted 3 lines · Xiaoyao Li

arch/x86/kernel/cpu/microcode/intel.c u32 intel_get_platform_id(void) · line 150

Quoted 1 line
arch/x86/kernel/cpu/microcode/intel.c +1 -0
+	/* Don't try to read microcode bits when virtualized. */

The platform ID is not only used by microcode update.

From the series with the blamed commit:
Today, the only code that cares about the platform ID is the microcode update code itself. To facilitate storing the platform ID in a more generic place and using it outside of the microcode update itself, put the enumeration into a helper function.

But it is still only used for microcode stuff AFAICT... It seems it was just moved to be "consistent with the other fields which can be matched".

Quoted 5 lines

I don't think
calling them microcode bits is proper. It's also stashed in struct cpuinfo_x86::intel_platform_id and used in x86_match_cpu() as a factor as the generic FMS (Family, Model, Stepping) bits.

Quoted 8 lines

We skip things due to hypervisor is set usually when the thing is known unable to be virtualized. From the perspective of microcode update, it's OK to skip reading it due to hypervisor bit. However, from the perspective of generic platform ID, it seems not that reasonable to skip reading it due to hypervisor bit. Especially KVM has supported this MSR for normal VMs.

So how about using the safe version of RDMSR, if we want the enhancement?

With the check if we are not virtualized we can get a nice warning if run on some weird platform with this MSR missing.

#14 lore raw

Re: [PATCH 2/2] x86/cpu: Skip reading MSR_IA32_PLATFORM_ID in virtualized environment

Quoted 8 lines · Rick P

From the series with the blamed commit:
Today, the only code that cares about the platform ID is the microcode update code itself. To facilitate storing the platform ID in a more generic place and using it outside of the microcode update itself, put the enumeration into a helper function.

But it is still only used for microcode stuff AFAICT... It seems it was just moved to be "consistent with the other fields which can be matched".

Right. The one place it is being queried is right next to the microcode version being queried. The microcode version is a field in that info structure, so the path of least resistance is to put the platform ID in there too. It also just happened to fit in an existing hole so it doesn't increase the size of the structure either.

#7 lore raw

Re: [PATCH 1/2] KVM: TDX: Allow TDs to read MSR_IA32_PLATFORM_ID

In reply to Sean Christopherson
Quoted 4 lines · Sean Christopherson
Quoted 2 lines

This is a problem with the bare metal KVM behavior too, and it's just super old?

No?  Oh, I see what you're asking.

I mean that KVM has hacked ABI around this already. So here we just don't do it again.

Quoted 4 lines

I'm mostly concerned about the host side of things.  The problem with TDX is that a TDX-Module update could effectively
change KVM's behavior, i.e. if the TDX-Module decides it needs to emulate PLATFORM_ID for whatever reason.

We have similar problems with CPUID bits.

Besides that, the guest can actually change some MSR handling. The reduced #VE thing.

Quoted 3 lines

  So not only would KVM need to enumerate to userspace that the MSR is supported/emulated for TDX guests, KVM would also need to differentiate between emulated by KVM and emulated by the TDX-Module.

The KVM supported and TDX-Module supported ones are already effectively separate things. And the guest can even know which one it is accessing.

I agree we can just not support this MSR for now, but I think Chao is right we need a plan for this. I don't have any clear and confident ideas.

I'm seeing a lot of parrallels to the CPUID bit cleanup effort. Binbin and the team did a bunch of pondering to find the right line between opt-ins everywhere and avoiding backwards ABI issues. We probably need to do this here too.

Quoted 10 lines
Quoted 9 lines

For TDX, hmm. I guess the standard thing to do in order to avoid creating a KVM ABI problems is just match the arch behavior. But for TDs, it is a very special type of VM. The special TDX guest things can't work on bare metal. Furthermore, guest opt ins can change what arch is even supposed to be virtualized. So the normal KVM default thing to do doesn't always fit.

So instead we will just virtualize as little as possible to keep Linux guest running? Ok.

Quoted 6 lines

Yeah, and the sequence of events matters.  Most of KVM's half-hearted emulation of random MSRs exists because KVM needed to be able to run existing kernels. But once KVM (and other hypervisors) existed, kernels learned to run as guests (and hardware vendors largely did a better job of explicitly enumerating MSRs and whatnot), and so the need to throw hacks into KVM mostly went away.

So now we don't want to throw hacks in again. But we (Intel TDX folks) need to make sure TDX arch makes it so there isn't a pressure for hacks too.

author
#9 lore raw

Re: [PATCH 1/2] KVM: TDX: Allow TDs to read MSR_IA32_PLATFORM_ID

In reply to Dave Hansen
Quoted 7 lines · Dave Hansen
Quoted 4 lines · Binbin Wu

Linux kernel reads MSR_IA32_PLATFORM_ID during init since commit d8630b67ca1e ("x86/cpu: Add platform ID to CPU info structure"). KVM already supports this MSR on read for normal VMs by returning 0. Without support for this MSR, TDs get unchecked MSR access errors.

NAK from me on this.

Based on the discussions, I will drop this patch.

Quoted 8 lines

The platform ID is used for one thing and one thing only: microcode updates. Those updates are solely the domain of the bare-metal OS.

The (guest) kernel code that's even trying to touch this MSR is buggy and insane. We need to turn that code it off when X86_FEATURE_HYPERVISOR==1. There's already a patch floating around to do that.

I guess it's the second patch in this patch set.

The second patch checks hypervisor bit via native_cpuid_ecx(1) instead of boot_cpu_has(X86_FEATURE_HYPERVISOR) since intel_get_platform_id() could be called in ucode load path even in virtualized environment when CONFIG_MICROCODE_DBG is set.

But using native_cpuid_ecx(1) will have an issue for XEN PV guest. intel_get_platform_id() can be called by XEN PV guest and the native_cpuid_ecx(1) doesn't have the hypervisor bit for XEN PV guest.

There could be two options:
- Use cpuid_ecx(1) in intel_get_platform_id() and drop the helper.
- Use boot_cpu_has(X86_FEATURE_HYPERVISOR) directly since CONFIG_MICROCODE_DBG is just a debug config option. Maybe add some comments to avoid confusion due to the pre-existing comments for intel_cpuid_vfm().

Quoted 3 lines

Please don't add more smoke and mirrors.

#15 lore raw

Re: [PATCH 2/2] x86/cpu: Skip reading MSR_IA32_PLATFORM_ID in virtualized environment

In reply to Binbin Wu

arch/x86/kernel/cpu/microcode/internal.h static inline unsigned int x86_cpuid_family(void) · line 106

Quoted 4 lines · Binbin Wu
arch/x86/kernel/cpu/microcode/internal.h +4 -0
+static inline bool x86_cpuid_has_hypervisor(void)
+{
+	return native_cpuid_ecx(1) & BIT(31);
+}

Could you please put X86_FEATURE_HYPERVISOR in here somewhere? Even if it's just (logically):

	return native_cpuid_ecx(1) & (X86_FEATURE_HYPERVISOR & 0x1f);